Deep Learning Based Method For Non-intrusive Residential Appliances Load Disaggregation

Energy conservation and emissions reduction are essential to promote economic development and improve environmental issues.The total electricity consumption in power entrance of residence can be decomposed to specific load electricity consumption through non-invasive residential electricity load decomposition method,and the operation state of each load can also be acquired.The decomposition results of which can be used to guide the using electricity behavior of customers,and be helpful to promote saving electricity and achieving energy conservation.However,The existing non-intrusive household load disaggregation algorithms are mainly based on the electrical characteristics of the power load to establish the load models,and use the optimization technology or pattern recognition technology to achieve the of the load disaggregation.Most of the researches above did not solve the classification of high-noise high-power and non-stationary load s.To solve the problems above,based on the latest progress of deep learning algorithms,a non-intrusive residential electricity load disaggregation method based on deep learning algorithm is proposed in this paper.Firstly,the total residential load power and the power of each electrical load are collected,the power data of the typical load in the house is expanded by using data expansion technology to generate enough data for training and improve the performance in deep learning models.After that,each load characteristic is trained and acquired through the model based on CNN and RNN which can automatically extract the load features and generate the relationship between mains reading and individual appliance energy usage.Then,according to the main reading information,power consumption of individual appliance can be estimated.Finally,the latest published UK-DALE public load data set in May 2016 is used as the experimental data set in this paper,and six evaluation metrics are adopted to evaluate the performance of the proposed load disaggregation models such as the disaggregation accuracy,proportion of total energy classified correctly,mean normalized error?recall?accuracy and F1-measure.The experimental results show that In general,the proposed five deep neural ne tworks have good load disaggregation performance and have good generalization performance.